基于四种人工智能模型的极端干旱区参考作物蒸散量研究

Study of Reference Crop Evapotranspiration in Extreme Arid Zone Based on Four Artificial Intelligence Models

摘要: 本文利用新疆维吾尔自治区阿克苏市气象站点的逐日气象数据，通过支持向量机(SVM)、高斯过程回归(GPR)、提升树(BT)以及BP神经网络(BPNN)四种人工智能算法模型，对1978~2018年极端干旱区的参考作物蒸散量(ET0)进行模拟，结果表明：1) 当输入气象因子相同时，SVM、GPR和BPNN的模型精度较高且较为接近，相比之下BT模型的精度稍差；四种人工智能模型当全部因子(T_max, T_min, R_mean, n, u₂, R_a)输入时模拟精度最好，其中GPR模型精度最高；当只输入4种气象因子时，各模型在组合4 (T_max, T_min, u₂, R_a)条件下的模拟精度较高，其中BPNN模型的模拟效果最好，为极端干旱区的模拟ET₀首选模型。2) 就季节尺度而言，BPNN模型对秋季ET₀模拟效果最好，而对夏季ET₀的模拟效果最差。3) 与传统经验公式Hargreaves、Irmark-Allen、Jensen-Haise、Makkink相比，人工智能模型在模拟ET₀时有明显的优势。

Abstract: In order to simulate the crop evapotranspiration (ET0) in extreme arid areas from 1978 to 2018, based on daily gauge observations from Aksu, Xinjiang Uygur Autonomous Region, four artificial intelligence algorithm models including Support Vector Machine (SVM), Gaussian Process Regression (GPR), Boosting Tree (BT), and BP Neural Network (BPNN) are established. The results suggest that: 1) Compared with the result of BT, when the input meteorological factors are the same, the SVM, GPR, and BPNN show the higher and closer accuracy, BT’s result is weaker than other models. The four models exhibit the highest accuracy when the input factors are complete (T_max, T_min, R_mean, n, u₂, R_a), and GPR is the best among them. When the number of the input factors is four, the accuracy under combination-4 (T_max, T_min, u₂, R_a) is superior, and the simulation effect of BPNN is the best, so it is the favored model to simulate the ET₀ in extreme arid areas. 2) On the seasonal scale, the BPNN demonstrates the highest simulation accuracy in autumn and the lowest accuracy occurs in summer. 3) Compared with the traditional empirical formulas involving Hargreaves, Irmark Allen, Jensen Haise, and Makkink, the artificial intelligence model has obvious advantages in simulating ET₀.